The study found that most people with suicidal thoughts post their views on social media and so the institute has developed a neural network which can identify over 80% of “at-risk tweets” on Twitter and other social media platforms by assessing the language used by the users.
Researchers from Midas Lab at IIIT-D have developed an advanced artificial intelligence which can identify words and sentences on social media that hint at suicide or suicidal thoughts.
Ramit Sawhney, who worked on the project, said initially they asked clinical psychology students to assess over 34,000 tweets and identify if they were suicidal in nature. “This was done to look for key words like ‘kill myself, ‘end my life’ etc. The students were asked to mark it as suicidal if they found the tweets as such. There was 88% agreement between students on what was suicidal,” Sawhney said.
The IIIT researcher informed that the same was fed into a neural network. “We used software that even Google uses in its search engine. The software was taught about language and how keywords can help find associated words.”
Professor Rajiv Ratan Shah of IIIT-D said that their research focused on building natural language processing system to identify potential suicidal intent in social media messages to provide support to “at-risk users” in an automated fashion without any human interference.
The Suicidality assessment Time-Aware Temporal Network (STATENet), developed by them, can identify over 80% of at-risk tweets on Twitter.
Shah said, “Recent studies showed that people exhibiting suicidal ideation make frequent use of social media to share their mental state, disclosing their suicidal thoughts and plans. This makes it essential to automatically flag at-risk posts to extend support.”
Sawhney said they are now looking to collaborate with University of California to improve mental health. “Currently Facebook and Twitter sends messages of suicide prevention helpline or other texts in case if the keywords written by a user explicitly suggest suicide. But it is not effective as it doesn’t identify the language as our neural network does. We are developing a model that can be picked up by social media companies,” he added.